From Data to Decisions: VLA Pipelines for Real Robots

Vision-Language-Action (VLA) models have emerged as a promising paradigm for building generally capable robot agents. Leveraging scalable teleoperation, simulation, and web-scale video data, as well as pre-trained vision and language models, these systems have demonstrated impressive generalization and instruction-following capability across diverse manipulation tasks.

However, their real-world impact remains limited: most pipelines lack rigorous data curation, principled training regimes, and standardized evaluation, leading to inconsistent performance and safety concerns on physical robots. Meanwhile, performance on real robots is dictated by the underlying "data engine", highlighting the importance of scalable data collection, domain randomization, procedural synthesis, and human-in-the-loop feedback.

This workshop aims to catalyze progress by examining the entire end-to-end VLA pipeline, from scalable data collection and systematic curation to training strategies and inference-time reasoning. A core focus will be on surfacing best practices and establishing shared metrics for generalization and safety.